Well....
The 3.5gb stride is all that can happen at once, that's 196GB/s. the segmented 500mb has the same bandwidth as every one of the other 7 memory blocks. They all have the same bandwidth capability, 28gb/s. Every last one of them
That's eight 500mb memory blocks at 28gb/s
28 X 8 = 224 GB/s
That's where we get the 224.
The segmented ram has the same speed as every other 500 mb block. Anandtech article actually states this clearly but soon afterwards they start making things really muddy. but so we can stay clear, let me restate-
The segmented ram is no slower than any other 500mb block when it is accesed.
Soon after anandtech article explains that, they start clumping all the other 500mb blocks together as one large 3.5gb memory block. Everyone is calling this the fast ram, but its actually seven 500mb blocks at 28GB/s. that's where you get the 196GB/sec figure.
But that's we're this entire thing is getting confused. If you were pushing every bit of all the ram thru at one time, it could only come from 7 of the 8 ram blocks in one large and organized stride. 7 blocks pushing 28GB/s all at once for a maximum of 196GB/s. if you were steady pushing 3.5gb worth of data, non stop over a span of time, the segmented 500mb would only hurt because it cannot be read while its neighboring block is being read at the same time......
BUT-
That's completely unlike any gaming load. The only way the scenario exist would be in a synthetic / benchmark condition that was created with the sole purpose of testing the entire bandwidth all at once. In a real life, a real gaming scenario, its just not a steady 3.5GB/s data push thru non stop over a long span of time. Your GPU would completely be overwhelmed, the GPU would be completely saturated. It's completely unrealistic and absolutely not how the GPU works in a normal gaming environment.
The truth is that there is no 3.5gb block of ram. It's 7 blocks of 500mb. And the data seldom flows in a massive 3.5gb stride, unless your benchmark is written to do that.
So, then we have the matter of the "slow" ram. The "slow" 500mb......
Well, it doesn't really work like that.
It's actually two 500mb blocks that share the same path. It's actually 1gb of ram that shares resources. The other 3gb can carry on all day long without any problems. But those 2blocks that share the 1 path, this is the compromise. Either data flows thru one segment, or data flows from the other. But not at the same time.
Putting the entire 3.5gb together and calling it the fast ram is just totally wrong. I mean, in a scenario that is pushing strictly bandwidth capacity, the data can flow all together as if it is one large stride but in the real world games aren't 3.5gb chunks that just pass thru non stop. Games have all sorts of stuff loaded into VRAM and ATM, no single item stored in vram is 3.5gb in size. That's crazy to talk like that.
But the truth is, as fas as gaming, its not the 500mb interfering with the 3.5gb. That will never be the case. It's a 500mb block sharing the same path as its neighboring 500mb. It's the last 1GB that is in struggle. This is why nvidia set the drivers to avoid trying to stuff anything in there, because then you have the XOR situation that can choak out the smm.
The segmented ram can be read along with the all the other 3gb. It can be read as fast as all the other 500mb blocks. But when it pushing dats, it blocks data flow from the 500mb block beside it. This is the issue, and why it is so interesting.
Sooner or later somewhere there needs to be better explanation of is.
Anandtech spoke of the all the other 7 memory blocks as if it was one 3.5gb block right after they explained that the segmented 500mb had the same transfer rate as all the other ram blocks. They went from talking about how it works to describing a specific scenario, one that is very unlikely in the real world.
Its probably not so much better when you consider that two 500mb blocks are competing for the same path, that this is an entire gb segment swapping priority to be read. But that's how it will be when you are gaming over 3500mb. Nvidia might find some magic way to deal with it, or they might not. I expect they are working on a driver to better manage this but have pulled down reference to try to save their butts legally when the lawsuits come in. If the driver comes out it will be carefully worded as not to be a "fix" for anything. Because they will stand by their stance that nothing is broke and it was intentional design. Their lawyers surely have been advising them and that is why they are scared to make any moves right now